Meeting Summary

  • The idea about use dynamic knowledge graph to build a persistent memory is theoretical feasible.
  • This topic is trending, which means many teams might working on it.
  • Such valuable long conversation dataset is hard to find, likely not open source but owned by large company.
  • Core question: what is the benchmark and evaluation metric?

Next Step

  • Do literature review deeply on these aspects to get insights about benchmark and current SOTA:
    • KG in conversational AI
    • RAG
    • AI memory system
    • User modeling and personalization
  • I should keep focus on:
    • Evaluation tasks for AI memory or long term interaction
    • Relevant datasets
    • The standard metric I can use
    • SOTA baselines